Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

RADIOMETRIC CALIBRATION OF VERY LARGE DIGITAL

AERIAL MOSAICS
Simon Collings, Peter Caccetta, Norm Campbell and Xiaoliang Wu
CSIRO Mathematical and Information Sciences
The Leeuwin Centre,
65 Brockway Rd, Floreat Park, WA
+61 8 9333 6690, +61 8 9333 6121
{Simon.Collings, Peter.Caccetta,Norm.Campbell, Xiaoliang.Wu }@csiro.au

Abstract
The advent of routine collection of high-quality digital photography provides for traditional uses
as well as “remote sensing” uses such as the monitoring of environmental indicators. A well-
devised urban monitoring system, based on consistent data and methods, provides the
opportunity to track and communicate changes in features of interest in a way that has not
previously been possible. Data that are geometrically and radiometrically consistent are
fundamental to establishing systems for monitoring. In this paper we focus on models for
radiometric calibration of mosaics consisting of thousands of images. We apply the models to
data acquired by CSIRO and its partners as part of regular, systematic acquisitions over the city
of Perth for a project known as Urban Monitor. One goal of the project, and hence the model
development, is to produce annually-updated mosaics calibrated to reflectance at 0.2m Ground
Sample Distance (GSD) for an area of approximately 9600sqkm. This equates to terabytes of
data and, for frame-based instruments, tens of thousands of images. For the experiments
considered in this paper, this requires mosaicking estimates derived from more than 30000
digital photographic frames. A key part of the processing is the removal of spectral variation due
to the viewing geometry, with this paper focusing on the effect typically attributed to the
Bidirectional Reflectance Distribution Function (BRDF) of the land surface. A variety of
techniques based on semi-empirical BRDF kernels have been proposed in the literature for
correcting the BRDF effect in single frames, but mosaics with many frames provide unique
challenges. This paper presents and illuminates a selection of kernel-based techniques for
BRDF correction of digital aerial frame mosaics, and incorporates combined additive and
multiplicative correction terms for correcting other effects, such as variations due to the sensor
and atmosphere. Using ground truth, which consists of lab-measured white, grey and black
targets that were placed in the field at the time of acquisition, we calculate the fundamental
limitations of each model, leading to an optimal result for each model type. We demonstrate
estimates of ground reflectance accurate to approximately 10%, 5% and 3% for ground targets
having a reflectance of 0.9, 0.4 and 0.04 respectively.

Introduction
Regular aerial photography acquisition programs are increasingly moving from
analogue to digital acquisition. For infrastructure and city planning applications,
cameras having high geometric and radiometric accuracy are typically used.
Regular acquisitions of data having these attributes make the data a potentially
useful source for developing monitoring systems.
In this paper, we will summarise the issues and challenges that must be
overcome in order to perform radiometric calibration (to reflectance) of digital
aerial frame data. In 2007, the Urban Monitor program acquired approximately
35000 frames of data over a period of 39 days beginning on the 14th of March
2007. We chose a sample of 3000 frames on which to perform our
experiments. Figure 1 shows the arrangement of the frames used in
experiments outlined herein, their dates of acquisition, along with the raw

1
Day
1
3
4
10
18
21
24
25
28
32
39

Figure 1: A) Arrangement and time of capture of frames in the Big Perth East dataset. Day
th
1 is the 14 of March 2007. B) Geocorrected raw mosaic.
geo-corrected images for this dataset, which is drawn from the 2007 Urban
Monitor mosaic. Observe that there is substantial variation in brightness, both
between and within the frames. The 2007 mosaic was acquired using a
Microsoft Ultra Cam D (Leberl and Gruber, 2003), capturing red, green, blue
and near-infra red, along with a panchromatic image.

2
Figure 2 Extract from the Urban Monitor 2007 mosaic, shows considerable variation in
intensity between frames.

The Bidirectional Reflectance Distribution Function (see e.g. Schaepman-Strub


et al. 2006 and references therein) describes the ratio of incident to reflected
light intensity as a function of the view and illumination angles. This angular
dependence is responsible for many of the spectral inconsistencies visible in
the mosaic above (the BRDF effect).
Currently-favoured methods for BRDF correction include the kernel-based
semi-empirical models of Wanner et al. (1995) and Roujean (1992), which are
fitted additively to the image data, and then used to subtract or multiply out the
BRDF effect. Recently these methods have been augmented for the purpose of
mosaicking, by including a penalty term to enforce similarity on the overlap (Wu
2006, Collings et al. 2008) and also to include spatial and spectral smoothness
(Collings et al. 2010 (1)).

3
It is desirable, if possible, to radiometrically calibrate the images to reflectance,
which is the ratio of reflected to incident flux, often expressed as a percentage.
To facilitate the calibration, the Urban Monitor project deploys targets each
year, consisting of fibreboard painted with matt finish white, grey or black paint.
These radiometric targets are distributed across the region to correspond with
the time of acquisition. The targets have reflectance measured in the laboratory
before being deployed in the field. The targets are placed in groups of three,
with one target of each colour placed next to each other in the scene, see
Figure 3. This arrangement exposes the strengths and weaknesses of the
various models used for BRDF correction.

BRDF Model
A widely used technique for the correction of the BRDF effect is to fit a model to
the image data and then use the model to remove the effect. The model can be
empirical (e.g. polynomials, Kennedy et al., 1997, Pickup et al. 1995), physical
(Roujean et al. 1992) or semi-empirical (Wanner et al. 1995), which combines
elements of both. In each case, the model is fitted to the image data, usually
with least squares regression, and then the image data is normalised to a
standard view and sun position by either subtracting or dividing out the trend
(see Kennedy et al. (1997) for details of how this is performed).
For our radiometric calibration technique, we have found it is necessary to
incorporate a gain term into the model, in order to account for differences in
aperture and exposure time of the camera from frame to frame. For pixel l of
image i in band j with intensity value ρ ijl , the corrected intensity value, ρ̂ ijl is
given by:
ρˆ ijl = (1 + α ij ) ρ ijl − β ij ( L(ξ il ) − Li ) − τ ij (2)

Where:
i = 1, K , N images
j = 1, K , B bands
θ il = solar zenith angle at pixel l , image i
ς il = view zenith angle
φil = relative azimuth angle between the solar and viewing directions
ξ il = (θ il , ς il , φil )
L = Li Sparse BRDF kernel
L i = the mean of the Li Sparse BRDF kernel
over the i th image
α ij = the gain parameter for image i band j
β ij = kernel parameter for image i band j
τ ij = the offset parameter for image i band j

4
We call this model the constant-gain-additive model because the gain term,
while varying from image to image, does not vary due to changing angles within
an image.

Let α = {α ij }i , j =1 β = {β ij }i , j =1 τ = {τ ij }i , j =1 be vectors of parameters. These are


N ,B N ,B N ,B

found by minimising an objective function that incorporates various constraints:


C1 (α , β ) + λ C 2 (α , β ,τ ) + γC 3 ( β ) + µC 4 (τ ) + ηC 5 (α ) (3)
Each constraint is weighted according to its relative importance, as controlled
by the parameters λ , γ , µ ,η > 0 . The terms of (3) are explained in detail in
Collings et al. 2011, and are summarised below:
C1 ( β ) is the internal cost function, which fits BRDF kernels to the data within
each image and attempts to ‘flatten’ the image intensity across its spatial
extents. It is based on the assumption that the image is uniform in colour and
that the corrected pixels’ intensities are normally distributed within any given
image.
C 2 (α , β ,τ ) is the overlap cost function (Wu 2006, Collings et al. 2008) which
penalises the sum-of-squares differences of corrected pixels in the overlap
between two images. It is based on the assumption that a scene point, once
corrected, should look the same from two different viewing positions.
C 3 ( β ) is the spatial smoothness cost function (see also Collings 2010 (2))
which penalises the square of the second spatial derivative of the kernel
coefficients. It is based on the assumption that neighbouring images should
have similar land-covers and hence their kernel coefficients should also be
similar.
C 4 (τ ) is the additive ridge cost function (see Hastie 2001) that penalises the
size of the offset parameters τ i . It is necessary because otherwise the equation
system will be singular as the offsets are only defined relative to each other.
C 5 (α ) is the multiplicative ridge cost function which is necessary to prevent the
multiplicative parameters from being set equal to -1, sending the corrected
values to 0.
The weights of C 4 (τ ) and C 5 (α ) control the absolute deviation that the
corrected values have from their original values. If there is a high degree of
confidence in the original values, these weights should be set to large numbers.
Expression (3) is minimised by differentiating with respect to α , β and τ and
setting the resulting set of equations to 0. This leads to a system of equations
for each band with 3 × N parameters (with N the number of images in the
mosaic). For the example shown in Figure 1 there are approximately 9000
parameters. The equation systems are sparse and banded and can be solve
with standard techniques (see, for example, Golub and van Loan, 1996).

5
Figure 3 Black grey and white targets placed within the mosaic. A) Pan
image. B) Multispectral image.
Radiometric Calibration Targets
The ground-truth data consist of fibreboard targets approximately 1.2m square
that have been painted white, grey or black. The target spectra were initially
measured in the laboratory with a spectrometer to obtain an estimate of
reflectance and then deployed into the field to be digitally photographed by the
camera during acquisition.
The targets were typically deployed in sets of three, consisting of a black, grey
and white target collocated in an area of interest, thus providing a low, medium
and high reflectance ground truth for calibration purposes, see Figure 3. The
size of the targets was chosen such that one image pixel, expressed as GSD,
would lie entirely within a given target, while at the same time considering the
operational practicalities of deploying targets using standard utility vehicles.
Larger targets could be constructed by forming arrays of these basic targets,
but were not used in the 2007 deployment in order to get targets placed over a
larger spatial extent.
The targets were all measured in the laboratory with an ASD FieldSpec Pro
spectrometer. Each target was measured in 4 different locations over its area
and the results were averaged and then convolved with the spectral responses
of the UltraCam D to produce a set of lab measurements comparable with the
field measurements.
The targets are then used in two ways to correct the mosaic to radiance. Firstly,
an additional constrain is incorporated into (3) that enforces similarity of the
corrected target pixel values (i.e. the target pixels are considered to be
overlapping). Secondly, the lab measurements of the targets are used to fit a

6
Colour Band Best Additive Practical
Model implementation
1 13.24 9.49
White 2 14.35 6.29
3 13.16 12.30
4 12.29 6.36
1 3.89 5.63
Grey 2 3.29 3.15
3 3.61 5.14
4 4.84 4.90
1 10.50 2.57
Black 2 11.69 1.37
3 10.62 2.01
4 9.38 3.65
Table 1: Shows the rms difference between the lab measured reflectance values of observations of
all targets and the corrected values (expressed as % of reflectance). The third column shows the
best theoretical additive model and the fourth column shows our practical implementation of the
constant gain additive model.

global offset and gain to the corrected mosaic so that the image target values
give the best agreement to the lab target values.

Best theoretical models


The collocation of the white, grey and black targets lead to a so-called “best
theoretical model” (Collings et al. 2010 (2)). The best theoretical model gives
the best result that any theoretical radiometric correction model can achieve at
correcting the target pixels to radiance.
Best theoretical models arise because the collocation of the targets means that,
in the absence of stratification, the correction applied to each target in the same
image must be approximately the same. Each colour of target optimally
requires a different correction. Therefore any practical correction must be a
compromise between correcting each of the coloured targets perfectly.
The above reasoning leads to a best theoretical model, which is an upper limit
on how well a particular radiometric calibration model can perform. In Table 1 a
comparison of the RMS difference between corrected target pixels and their lab
measurements is shown. The third column shows the best additive model for
this mosaic and target arrangement. The fourth column shows how our
constant-gain-additive model performs.
For a more detailed description of best theoretical models, the reader is
referred to Collings et al. 2010 (2). The model presented here, which includes a
gain term, outperforms the best theoretical additive model in correcting targets
to reflectance, and so can be said to be superior in this function than any
possible additive model.

7
Figure 4 Extract from the Urban Monitor 2007 mosaic after calibration shows that the
between frame spectral inconsistencies have been greatly reduced.

Results and Conclusion


Figure 4 shows the corrected extract of the Big Perth East mosaic from Figure
3. It is clear that many of the inconsistencies between the frames comprising
the mosaic have been removed. Table 1 shows that the constant-gain-additive
model outperforms the best additive model, which means that no practical
implementation of an additive model can give better results.
In this paper we have outlined some techniques applicable for radiometric
calibration of very large aerial frame mosaics. By constraining the process of
fitting BRDF kernels, a robust result can be obtained. Incorporating a
multiplicative parameter into the calibration procedure means that differences in
aperture and exposure time for individual frames can be accounted for. This
means that a superior result can be obtained for correcting targets to radiance.
Indeed the practical implementation of the method has out-performed the best

8
theoretical additive model in RMS difference from corrected target pixels
compared to lab measurements.
References
Collings, S., Wu, X., Caccetta P. and Campbell, N., 2008. Fitting BRDF kernels
with spatial smoothness and overlap constraints. Proceedings of the 14th
Australasian Remote Sensing and Photogrammetry Conference, Darwin,
Australia.
Collings, S., X. Wu, P. Caccetta and N. Campbell. Techniques for BRDF
correction of hyperspectral mosaics. IEEE Transacions on Geoscience and
Remote Sensing 48;10 (2010).
Collings, S., X. Wu, P. Caccetta and N. Campbell. Empirical models for
radiometric calibration of frame mosaics. To appear in IEEE Transacions on
Geoscience and Remote Sensing (Under revision August 2010).
Golub, G.H. and Van Loan, C.F. 1996. Matrix Computations. Johns Hopkins
University Press. 3rd Edition.
Hastie, T., Tibshirani, R. and Friedman, J., 2001. The Elements of Statistical
Learning. Springer.

Kennedy, R. E., Cohen, W. B. and Takao, G., 1997. Empirical methods to


compensate for a view-angle dependent brightness gradient in aviris imagery.
Remote Sensing of Environment, (62)277-291.
Leberl, F.; Gruber, M., 2003. Flying the new large format digital aerial camera
Ultracam. In Photogrammetric Week ´03; Fritsch, D., Ed.; Wichmann Verlag:
Heidelberg, Germany; pp. 67-76.
Pickup, G., Chewings, V.H. and Pearce, G., 1995. Procedures for correcting
high resolution airborne video imagery. International Journal of Remote
Sensing, 16(9):1647-1667.
Roujean, J. L., Leroy, M. and Deshchamps, P. Y., 1992. A bidirectional
reflectance model of the earth's surface for the correction of remote sensor
data. Journal of Geophysical Research, 97(D18):455-468.
Schaepman-Strub, G., Scaepman, M.E., Painter, T.H., Dangel, S., and
Martonchik, J.V., 2006. Reflectance quantities in remote sensing - definitions
and case studies. Remote Sensing of Environment, (103):27-42.
Wanner, W., Li, X. and A.H. Strahler. On the derivation of kernels for kernel
driven models of bidirectional reflectance. Journal of Geophysical Research,
100:21077-21089, 1995.
th
Wu, X., 2006. Radiometric calibration of digital aerial imagery. The 13
Australasian Remote Sensing and Photogrammetry Conference Proceedings,
Canberra, Australia 2006.

You might also like